kibetbrian74/Multi-Scale-Driver-Interactions-on-Crop-Phenology-via-Graph-Based-Ensembled-Machine-Learning

It majorly involves quantifying climate-soil-topography interactions in the Muringato Catchment Area, Nyeri County, Kenya. However, it can be scaled to other areas. You just change the study area and run the code. I hope you know how the area of study is changed. If not, hit me up # Multi-Scale Driver Interactions on Crop Phenology via Graph-Based Ensembled Machine Learning #### ⚠ If you get `Unable to render code block` error when opening any of the scripts, refresh the page. It will work ## Project Overview This project investigates how climate, soil, and topographic factors interact across multiple spatial and temporal scales to influence crop phenology. We leverage satellite-derived vegetation indices (MODIS NDVI/EVI) as proxies for crop greenness, together with climate and terrain data, to model the timing of phenological events (e.g. start/end of growing season). Vegetation phenology is known to be primarily controlled by climatic drivers (precipitation, temperature) with terrain modulating local variability. To handle large geospatial datasets, we use Google Earth Engine-a cloud platform for planetary-scale satellite data analysis. Our novel approach builds a graph-based ensemble ML framework combining a Graph Convolutional Network (GCN) to capture non-linear interactions among drivers and an XGBoost model for robust regression. ## Objectives Main Aim: Develop a graph-based ensemble ML framework to quantify and predict multi-scale interactions of dynamic climate, static soil, and topographic drivers on crop phenological stages. Data Integration: Acquire, preprocess, and harmonize multi-source geospatial datasets (climate time series, soil maps, DEM/topography, vegetation indices) and engineer features that capture key cross-scale interactions. Graph & EDA: Construct graph representations of driver interactions and perform exploratory data analysis (EDA) to identify synergies, correlations, and spatial/temporal patterns influencing phenology. Modeling: Implement and optimize an ensemble of two models-a graph-oriented model (GCN) and a complementary model (XGBoost)-to predict phenological metrics, using spatial-temporal cross-validation for evaluation. Interpretation: Apply explainable AI techniques (e.g. SHAP values, Accumulated …